How to compare dexterous hands beyond the DOF number
Build a fair comparison of hand motion, tactile sensing, weight and software. Use a shared task and evidence instead of treating one specification as a ranking.
Build an evidence-based comparison that preserves actuation, mass and sensing definitions.
Compare a task, then a specification
Start with an object and a desired change: holding a cup, rotating a pen or inserting a card are different tasks. Define the starting pose, allowed time, success rule and permitted human help. A high joint count cannot by itself tell you whether the thumb reaches the required contact or whether the controller can regulate a delicate grasp. Use specifications to shortlist candidates and a common task to evaluate them.
Active DOF, joints and fingers are different counts
A joint count describes mechanical structure. Independent motion coordinates describe how that structure can move, while the available actuators and control modes determine what can be commanded. Coupled or compliant joints can move without separate motors. Preserve each publisher’s definition. LEAP v2 Standard explicitly distinguishes its powered motors from its larger DOF count; Allegro V4 describes independently controlled joints. Those architectures should not be compared using an unlabeled number.
Treat tactile sensing as an interface
Ask what the sensor actually reports, where it is placed and how it is calibrated. Contact detection, a pressure distribution and an optical tactile image are different observations. Check timestamps, rate, drift and software access. For a slip-sensitive task, define an experiment that measures whether feedback changes the outcome. Sensor count alone does not establish that the hand can use the signal effectively in a closed loop.
Keep weight and load boundaries visible
Compare one complete hand with one complete hand, and state whether a forearm, controller or wrist adapter is included. A palm-only mass should not be substituted for the total load on a robot arm. Distinguish fingertip force, hook lifting, a particular grasp payload and continuous operation. The object geometry and contact posture can change the result. Retain source qualifications rather than combining impressive numbers from different tests.
Inspect the developer path
Find the exact hardware revision in the SDK documentation, robot model and research repository. Check joint naming, supported control modes, installation instructions and license terms. A paper using a related generation is useful context but does not prove compatibility with your unit. Before buying, ask for a small, documented example that matches the signals your project needs. List any missing asset as an open integration question.
Make the final comparison explainable
Create three columns for every important requirement: public claim, evidence available and the check still needed. Prefer clear unknowns to an unsupported score. Ask candidates to perform the same task and report all attempts, assistance and failures. Keep the hand revision, software and sensing configuration beside the results. This produces a decision you can revisit when a new release or a new task changes the requirements.
A fair comparison starts with the same question
| Requirement | Evidence to compare | Common mistake |
|---|---|---|
| Finger motion | Independent controls, coupling and reachable contacts | Ranking by total DOF alone |
| Touch | Signal type, calibration and feedback access | Equating all sensors |
| Arm integration | Complete installed mass and mounting | Comparing palm-only to system mass |
| Manipulation | Matched task trials and assistance | Treating a selected clip as a benchmark |
| Developer access | Exact SDK/model release and license | Assuming a family link fits every model |
Try it yourself
FINGER MOTION LAB · two-joint approximation
One finger, two joints. How many independent motions?
2 joints · 2 independent coordinates
This diagram isolates two joints of one finger; it is not a whole hand. The coupled mode imposes an exact constraint. Real underactuated hands may have additional passive or compliant motion. This diagram does not model contact, torque or any commercial hand.
Your practical task
- Define one task and its acceptance condition.
- Pick two hands and preserve each specification’s scope.
- Mark public claims, supporting evidence and unknowns separately.
What to produce: A two-hand evidence sheet and a common validation task.
Check your understanding
Original sources & next steps
LEAP v2: coupling and powered motors ↗Allegro V4: model specifications ↗MuJoCo: coordinates, joints and actuation ↗Original Dexhands teaching material. Lesson and linked references reviewed 2026-09-24. Research links are not endorsements or evidence of hardware compatibility.